idnits 2.17.1 draft-falk-xcp-spec-02.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- ** It looks like you're using RFC 3978 boilerplate. You should update this to the boilerplate described in the IETF Trust License Policy document (see https://trustee.ietf.org/license-info), which is required now. -- Found old boilerplate from RFC 3978, Section 5.1 on line 17. -- Found old boilerplate from RFC 3978, Section 5.5 on line 1482. -- Found old boilerplate from RFC 3979, Section 5, paragraph 1 on line 1459. -- Found old boilerplate from RFC 3979, Section 5, paragraph 2 on line 1466. -- Found old boilerplate from RFC 3979, Section 5, paragraph 3 on line 1472. ** This document has an original RFC 3978 Section 5.4 Copyright Line, instead of the newer IETF Trust Copyright according to RFC 4748. ** This document has an original RFC 3978 Section 5.5 Disclaimer, instead of the newer disclaimer which includes the IETF Trust according to RFC 4748. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- == No 'Intended status' indicated for this document; assuming Proposed Standard Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack separate sections for Informative/Normative References. All references will be assumed normative when checking for downward references. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the RFC 3978 Section 5.4 Copyright Line does not match the current year == Using lowercase 'not' together with uppercase 'MUST', 'SHALL', 'SHOULD', or 'RECOMMENDED' is not an accepted usage according to RFC 2119. Please use uppercase 'NOT' together with RFC 2119 keywords (if that is what you mean). Found 'MUST not' in this paragraph: The X, Delta_Throughput, and RTT fields are unused and SHOULD be set to zero. A router MUST not perform any processing on a minimal format header. This format is intended for use in empty ACK packets, to return congestion information from receiver to sender. -- The document seems to lack a disclaimer for pre-RFC5378 work, but may have content which was first submitted before 10 November 2008. If you have contacted all the original authors and they are all willing to grant the BCP78 rights to the IETF Trust, then this is fine, and you can ignore this comment. If not, you may need to add the pre-RFC5378 disclaimer. (See the Legal Provisions document at https://trustee.ietf.org/license-info for more information.) -- The document date (November 5, 2006) is 6374 days in the past. Is this intentional? Checking references for intended status: Proposed Standard ---------------------------------------------------------------------------- (See RFCs 3967 and 4897 for information about using normative references to lower-maturity documents in RFCs) == Missing Reference: 'TBD' is mentioned on line 342, but not defined == Missing Reference: 'IP2' is mentioned on line 497, but not defined == Missing Reference: 'XCP' is mentioned on line 500, but not defined == Missing Reference: 'IP1' is mentioned on line 499, but not defined == Missing Reference: 'TCP' is mentioned on line 501, but not defined == Unused Reference: 'Padhye98' is defined on line 1354, but no explicit reference was found in the text == Outdated reference: A later version (-13) exists of draft-ietf-dccp-spec-04 -- Possible downref: Non-RFC (?) normative reference: ref. 'Jacobson88' -- Possible downref: Non-RFC (?) normative reference: ref. 'KHR02' -- Possible downref: Non-RFC (?) normative reference: ref. 'Kapoor05' -- Possible downref: Non-RFC (?) normative reference: ref. 'Katabi03' -- Possible downref: Non-RFC (?) normative reference: ref. 'Katabi04' -- Possible downref: Non-RFC (?) normative reference: ref. 'Padhye98' ** Obsolete normative reference: RFC 793 (Obsoleted by RFC 9293) ** Obsolete normative reference: RFC 813 (Obsoleted by RFC 7805) ** Obsolete normative reference: RFC 2309 (Obsoleted by RFC 7567) ** Obsolete normative reference: RFC 2401 (Obsoleted by RFC 4301) ** Obsolete normative reference: RFC 2402 (Obsoleted by RFC 4302, RFC 4305) ** Obsolete normative reference: RFC 2406 (Obsoleted by RFC 4303, RFC 4305) ** Obsolete normative reference: RFC 2581 (Obsoleted by RFC 5681) ** Obsolete normative reference: RFC 2960 (Obsoleted by RFC 4960) ** Downref: Normative reference to an Informational RFC: RFC 3135 -- Possible downref: Non-RFC (?) normative reference: ref. 'Zhang05' Summary: 13 errors (**), 0 flaws (~~), 10 warnings (==), 14 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group A. Falk 3 Internet-Draft Y. Pryadkin 4 Expires: May 9, 2007 ISI 5 D. Katabi 6 MIT 7 November 5, 2006 9 Specification for the Explicit Control Protocol (XCP) 10 draft-falk-xcp-spec-02.txt 12 Status of this Memo 14 By submitting this Internet-Draft, each author represents that any 15 applicable patent or other IPR claims of which he or she is aware 16 have been or will be disclosed, and any of which he or she becomes 17 aware will be disclosed, in accordance with Section 6 of BCP 79. 19 Internet-Drafts are working documents of the Internet Engineering 20 Task Force (IETF), its areas, and its working groups. Note that 21 other groups may also distribute working documents as Internet- 22 Drafts. 24 Internet-Drafts are draft documents valid for a maximum of six months 25 and may be updated, replaced, or obsoleted by other documents at any 26 time. It is inappropriate to use Internet-Drafts as reference 27 material or to cite them other than as "work in progress." 29 The list of current Internet-Drafts can be accessed at 30 http://www.ietf.org/ietf/1id-abstracts.txt. 32 The list of Internet-Draft Shadow Directories can be accessed at 33 http://www.ietf.org/shadow.html. 35 This Internet-Draft will expire on May 9, 2007. 37 Copyright Notice 39 Copyright (C) The Internet Society (2006). 41 Abstract 43 This document contains an initial specification for the Explicit 44 Control Protocol (XCP), an experimental congestion control protocol. 45 XCP is designed to deliver the highest possible end-to-end throughput 46 over a broad range of network infrastructure, including links with 47 very large bandwidth-delay products, which are not well served by the 48 current control algorithms. XCP is potentially applicable to any 49 transport protocol, although initial testing has applied it to TCP in 50 particular. XCP routers are required to perform a small calculation 51 on congestion state carried in each data packet. XCP routers also 52 periodically recalculate the local parameters required to provide 53 fairness. On the other hand, there is no per-flow congestion state 54 in XCP routers. 56 This version specification (-02) includes protocol changes that move 57 per-packet divisions from the router to the sender. 59 Table of Contents 61 1. Changes Since Last Version . . . . . . . . . . . . . . . . . . 3 62 2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 4 63 2.1. XCP Protocol Overview . . . . . . . . . . . . . . . . . . 5 64 3. The Congestion Header . . . . . . . . . . . . . . . . . . . . 9 65 3.1. Header placement . . . . . . . . . . . . . . . . . . . . . 9 66 3.2. Congestion Header Formats . . . . . . . . . . . . . . . . 9 67 3.3. IPsec issues . . . . . . . . . . . . . . . . . . . . . . . 12 68 3.4. NAT, middlebox issues . . . . . . . . . . . . . . . . . . 12 69 3.5. MPLS/Tunneling Issues . . . . . . . . . . . . . . . . . . 13 70 4. XCP Functions . . . . . . . . . . . . . . . . . . . . . . . . 14 71 4.1. End-System Functions . . . . . . . . . . . . . . . . . . . 14 72 4.1.1. Sending Packets . . . . . . . . . . . . . . . . . . . 14 73 4.1.2. Processing Feedback at the Receiver . . . . . . . . . 16 74 4.1.3. Processing Feedback at the Sender . . . . . . . . . . 16 75 4.2. Router functions . . . . . . . . . . . . . . . . . . . . . 18 76 4.2.1. Calculations Upon Packet Arrival . . . . . . . . . . . 19 77 4.2.2. Calculations Upon Control Interval Timeout . . . . . . 20 78 4.2.3. Calculations Upon Packet Departure . . . . . . . . . . 22 79 4.2.4. The Control Interval . . . . . . . . . . . . . . . . . 25 80 4.2.5. Obtaining the Persistent Queue . . . . . . . . . . . . 25 81 5. Unresolved Issues . . . . . . . . . . . . . . . . . . . . . . 28 82 5.1. XCP With Non-XCP Routers . . . . . . . . . . . . . . . . . 28 83 5.2. Variable Rate Links . . . . . . . . . . . . . . . . . . . 29 84 5.3. XCP as a TCP PEP . . . . . . . . . . . . . . . . . . . . . 29 85 5.4. Sharing resources between XCP and TCP . . . . . . . . . . 30 86 5.5. A Generalized Router Model . . . . . . . . . . . . . . . . 30 87 5.6. Host back-to-back operation . . . . . . . . . . . . . . . 30 88 6. Security Considerations . . . . . . . . . . . . . . . . . . . 32 89 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 34 90 8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 35 91 9. References . . . . . . . . . . . . . . . . . . . . . . . . . . 35 92 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 38 93 Intellectual Property and Copyright Statements . . . . . . . . . . 39 95 1. Changes Since Last Version 97 Changes between version -01 and -02 99 o Minor edits and typo fixes. 101 Changes between version -00 and -01 103 o Updated protocol to reflect movement of the per-packet division 104 from the router to the end-system. 106 o Incremented version number (to 0x02) to reflect change in packet 107 header format. 109 o Reordered the Protocol, Length, Version, and Format fields in the 110 congestion header (in anticipation of future support of IPv6 111 extension headers). 113 o Routers now MUST (from SHOULD) ignore fields other than 114 reverse_feedback when minimal header is used. 116 o No longer ignore packets with RTT set to zero. Senders with 117 coarse-grained timer may generate these if the RTT is less than 118 the timer precision. 120 o Added 'Open Issues' section on variable-rate links. 122 2. Introduction 124 The Van Jacobson congestion control algorithms [Jacobson88] [RFC2581] 125 are used by the Internet transport protocols TCP [RFC0793] and SCTP 126 [RFC2960]. The Jacobson algorithms are fundamental to stable and 127 efficient Internet operation, and they have been highly successful 128 over many orders of magnitude of Internet bandwidth and delay. 130 However, the Jacobson congestion control algorithms have begun to 131 reach their limits. Gigabit-per-second file transfers, lossy 132 wireless links, and high latency connections are all driving current 133 TCP congestion control outside of its natural operating regime. The 134 resulting performance problems are of great concern for important 135 network applications. 137 The original Jacobson algorithm was a purely end-to-end solution, 138 requiring no congestion-related state in routers. More recent 139 modifications have backed off from this purity. Active queue 140 management (AQM) in routers (e.g., RED) [RFC2309] improves 141 performance by keeping queues small, while Explicit Congestion 142 Notification (ECN) [RFC3168] passes one bit of congestion information 143 back to senders. These measures do improve performance, but there is 144 a limit to how much can be accomplished without more information from 145 routers. The requirement of extreme scalability together with 146 robustness has been a difficult hurdle to accelerating information 147 flow. 149 This document concerns the Explicit Control Protocol (XCP) developed 150 by Dina Katabi of MIT [KHR02]. XCP represents a significant advance 151 in Internet congestion control: it extracts congestion information 152 directly from routers, without any per-flow state. XCP should be 153 able to deliver the highest possible application performance over a 154 broad range of network infrastructure, including extremely high speed 155 and very high delay links that are not well served by the current 156 control algorithms. XCP achieves fairness, maximum link utilization, 157 and efficient use of bandwidth. XCP is novel in separating the 158 efficiency and fairness policies of congestion control, enabling 159 routers to put available capacity to work quickly while 160 conservatively managing the allocation of capacity to flows. XCP is 161 potentially applicable to any transport protocol, although initial 162 testing has applied it to TCP in particular. 164 XCP's scalability is built upon the principle of carrying per-flow 165 congestion state in packets. XCP packets carry a congestion header 166 through which the sender requests a desired throughput. Routers make 167 a fair per-flow bandwidth allocation without maintaining any per-flow 168 state. This enables the sender to learn the bottleneck router's 169 allocation to a particular flow in a single round trip. 171 The gains of XCP come with some pain. XCP is more difficult to 172 deploy than other proposed Internet congestion control improvements, 173 since it requires changes in the routers as well as in end systems. 174 It will be necessary to develop and test XCP with real user traffic 175 and in real environments, to gain experience with real router and 176 host implementations and to collect data on performance. Providing 177 specifications is an important step towards enabling experimentation 178 which, in turn, will lead to deployment. XCP deployment issues will 179 be addressed in more detail in a subsequent version of this document. 181 This document contains an initial specification of the protocol and 182 algorithms used by XCP, as an experimental protocol. The XCP 183 algorithms defined here are based upon Katabi's SIGCOMM paper 184 [KHR02], her MIT thesis [Katabi03], and her ns simulation. However, 185 this document includes algorithmic modifications and clarifications 186 that have arisen from early experience with implementing and testing 187 XCP at the USC Information Sciences Institute. (See 188 http://www.isi.edu/isi-xcp for our project page.) This document is 189 intended to provide a baseline for further engineering and testing of 190 XCP. 192 This document is organized as follows. The remainder of Section 2 193 provides an overview of the XCP protocol, Section 3 discusses the 194 format of the congestion header, Section 4 describes the functions 195 occurring in the end-systems and routers, and Section 5 lists some 196 unresolved issues. 198 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 199 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 200 document are to be interpreted as described in [RFC2119]. 202 2.1. XCP Protocol Overview 204 The participants in the XCP protocol include sender hosts, receiver 205 hosts, and intermediate nodes in which queuing occurs along the path 206 from the sender to the receiver. The intermediate nodes are 207 generally routers, but link-layer switches may also contain packet 208 queues. 210 XCP supplies feedback from the network to the sender on the maximum 211 rate (throughput) for injecting data into the network. XCP feedback 212 is acquired through the use of a congestion header on each packet 213 that is sent. Routers along the path may update the congestion 214 header as it moves from the sender to the receiver. The receiver 215 copies the network feedback into outbound packets of the same flow. 216 An end-system may function as both a sender and a receiver in the 217 case of a bidirectional flow. 219 The figure below illustrates four entities participating in XCP. The 220 sender initializes the congestion header, two routers along the way 221 may update it, and the receiver copies the feedback from the network 222 into a returning packet in the same flow. 224 +----------+ +--------+ +--------+ +----------+ 225 | |------->| Router |---->| Router |------->| | 226 | Sender | +--------+ +--------+ | Receiver | 227 | |<----------------------------------------| | 228 +----------+ +----------+ 230 The congestion header contains four pieces of data: 232 o RTT: Set by the sender to its current estimate of the round-trip 233 time. 235 o X: Set by the sender to its current estimate of inter-packet time 236 gap. This quantity is used in place of Throughput in earlier 237 drafts of these specifications to avoid per-packet division in the 238 router. See Section 4.1.1 for more about X. 240 o Delta_Throughput: Initialized to the amount which the sender would 241 like to change (increase or decrease) its throughput, and updated 242 by the routers along the path to be the network's allocated change 243 in throughput. This value will be a negative number if an XCP- 244 capable queue along the path wants the sender to slow down. 246 o Reverse_Feedback: When a data packet reaches the receiver, its 247 Delta_Throughput value is returned to the sender in the 248 Reverse_Feedback field of a congestion header of a returning 249 packet (e.g., in an ACK packet). 251 An XCP-capable router calculates a fair capacity re-allocation for 252 each packet. A flow only receives this re-allocation from a 253 particular router if that router is the bottleneck for that flow. 254 For XCP, a bottleneck router is defined to be a router that has 255 insufficient capacity to accept a flow's current or desired 256 throughput. 258 Based on current conditions, an XCP-capable router generates positive 259 or negative feedback each time a packet arrives and compares it the 260 packet's Delta_Throughput field. Delta_Throughput is reduced if the 261 current value exceeds this calculated feedback allocation. Each XCP- 262 capable router along the path from sender to receiver performs this 263 processing. A packet reaching the receiver therefore contains the 264 minimal feedback allocation from the network, i.e., the capacity 265 reallocation from the bottleneck router. 267 The receiver copies this value into the Reverse_Feedback field of a 268 returning packet in the same flow (e.g., an ACK or DATA-ACK for TCP) 269 and, in one round-trip, the sender learns the flow's per-packet 270 throughput allocation. 272 The sender uses the reverse feedback information to adjust its 273 allowed sending rate. For the transport protocol TCP [RFC0793], for 274 example, this may be accomplished by adjusting the congestion window, 275 or cwnd, that limits the amount of unacknowledged data in the 276 network. (Cwnd is defined for Van Jacobson congestion control in 277 [RFC2581].) 279 Additionally, it is possible to use XCP's explicit notification of 280 the bottleneck capacity allocation for other types of applications. 281 For example, XCP may be implemented to support multimedia streams 282 over DCCP [I-D.ietf-dccp-spec] or other transport protocols. 284 An XCP-capable router maintains two control algorithms on each output 285 port: a congestion controller and a fairness controller. The 286 congestion controller is responsible for making maximal use of the 287 outbound link while at the same time draining any standing queues. 288 The fairness controller is responsible for fairly allocating 289 bandwidth to flows sharing the link. These two algorithms are 290 executed only periodically, at an interval known as the "control 291 interval". The algorithms defined below set this interval to the RTT 292 averaged across all flows. Further work on choosing an appropriate 293 value for the control interval may be required. 295 Each port-specific instance of XCP is independent of every other, and 296 references to an "XCP router" should be considered an instance of XCP 297 running on a particular output port. 299 Actually, it is an oversimplification to say that congestion in 300 routers only appears at output ports. Routers are complex devices 301 which may experience resource contention in many forms and 302 locations. Correctly expressing congestion which doesn't occur at 303 the router output port is a topic for further study. Even so, it 304 is important to correctly identify where the queue will build up 305 in a router. The XCP algorithm will drain a standing queue; 306 however it is necessary to measure that queue in order for correct 307 operation. For more discussion of this issue see Section 308 Section 5.5. 310 More context, analysis, and background can be found in [KHR02] and 311 [Katabi03]. 313 3. The Congestion Header 315 The congestion control data required for XCP are placed in a new 316 header which is called the Congestion Header. 318 3.1. Header placement 320 The Congestion Header is located between the IP and transport 321 headers. This is consistent with the fact that XCP is neither hop- 322 by-hop communication -- as in IP -- nor end-to-end communication -- 323 as in TCP or UDP -- but is rather end-system-to-network 324 communication. It should allow a router to "easily" locate the 325 congestion header on a packet with no IP options. 327 Other choices were considered for header location. For example, 328 making the Congestion Header a TCP option was suggested. This 329 made sense as the congestion information is related to the 330 transport protocol. However, it requires that routers be aware of 331 the header format for every new transport protocol that might ever 332 use XCP. This seemed like an unreasonable burden to place on the 333 routers and would impede deployment of new transport protocols 334 and/or XCP. 336 It has also been suggested that the Congestion Header be an IPv4- 337 style option. While this proposal is transport protocol 338 independent, it would generally force XCP packets to take the slow 339 path on non-XCP routers along a path. This could severely impact 340 performance. 342 The XCP protocol uses protocol number [TBD], assigned by IANA. IP 343 packets containing XCP headers will use this protocol number in the 344 IP header's Protocol field [RFC0791] to indicate to routers and end- 345 systems that an XCP congestion header follows the IP header. 347 3.2. Congestion Header Formats 349 This section defines the XCP Congestion Header formats. This holds 350 for IPv4; the corresponding header for IPv6 is a subject for further 351 study. 353 XCP-capable IPv4 packets carry the following Congestion Header: 355 0 1 2 3 356 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 357 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 358 | Protocol | Length |Version|Format | unused | 359 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 360 | RTT | 361 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 362 | X | 363 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 364 | Delta_Throughput | 365 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 366 | Reverse_Feedback | 367 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 369 Protocol: 8 bits 371 This field indicates the next-level protocol used in the data 372 portion of the packet. The values for various protocols are 373 specified by IANA. 375 Length: 8 bits 377 This field indicates the length of the congestion header, measured 378 in bytes. Length in this version of XCP will always be 20 bytes, 379 or 0x14. 381 Version: 4 bits 383 This field indicates the version of XCP that is in use. The 384 version of XCP described in this document corresponds to a value 385 of 0x02. Future values will be assigned by IANA 386 (http://www.iana.org). See note in the IANA Considerations 387 section. 389 Format: 4 bits 391 This field contains a code to indicate the congestion header 392 format. The current format codes are defined below. 394 +-----------------+------+ 395 | Format | Code | 396 +-----------------+------+ 397 | Standard Format | 0x1 | 398 | | | 399 | Minimal Format | 0x2 | 400 +-----------------+------+ 402 Table 1 404 Standard Format 406 The standard format includes the X, Delta_Throughput, and RTT 407 fields shown above. This format is used by XCP in data packets 408 flowing from sender to receiver. 410 Minimal Format 412 The X, Delta_Throughput, and RTT fields are unused and SHOULD be 413 set to zero. A router MUST not perform any processing on a 414 minimal format header. This format is intended for use in empty 415 ACK packets, to return congestion information from receiver to 416 sender. 418 Other formats may be defined in the future, to define different 419 representation formats for the X, Delta_Throughput, and/or RTT 420 fields, for example. The corresponding format values format values 421 will be assigned by IANA. See IANA Considerations section below. 423 unused: 8 bits 425 This field is unused and MUST be set to zero in this version of 426 XCP. 428 RTT: 32bits 430 This field indicates the round-trip time measured by the sender, 431 in fixed point format with 28 bits after the binary point, in 432 seconds. Thus, the value of 1 corresponds to 2^(-28) of a second. 433 This field is an unsigned integer. 435 The minimum value expressible in this field is 0s. A value of 436 zero in the RTT field is legal and indicates that the sender 437 either does not yet know the round-trip time, or operates at a 438 coarse-grained timer granularity. The maximum value expressible 439 in this field is 15.9999999963 seconds, in steps of 3.7ns. 441 X: 32 bits 443 This field indicates the inter-packet time of the flow as 444 calculated by the sender, in fixed point format with 28 bits after 445 the binary point, in seconds. This is the same format as used by 446 the RTT field. 448 Delta_Throughput: 32 bits 450 This field indicates the desired or allocated change in 451 throughput. It is set by the sender to indicate the amount by 452 which the sender would like to adjust its throughput, and it may 453 be subsequently reduced by routers along the path (See 454 Section 4.2). It is measured in bytes per second and is a signed, 455 2's complement value. 457 The minimum throughput change expressible in this field is -17 458 Gbps. The maximum value expressible in this field is 17 Gbps, in 459 steps of 8 bits per second. 461 Reverse_Feedback: 32bits 463 This field indicates the value of Delta_Throughput received by the 464 data receiver. The receiver copies the field Delta_Throughput 465 into the Reverse_Feedback field of the next outgoing packet in the 466 same flow. See Section 4.1.2. 468 3.3. IPsec issues 470 IPsec [RFC2401] must be slightly modified to accommodate use of XCP. 471 The specifications for the IP Authenticated Header (AH) [RFC2402] and 472 IP Encapsulating Security Payload (ESP) [RFC2406] state that the 473 IPsec headers immediately follow the IP header. This would be a 474 problem for XCP in that a) it would make the XCP headers harder to 475 find by the routers, b) ESP encryption would make it impossible for 476 routers along the path to read and write congestion header 477 information and c) AH authentication would fail if any router along 478 the path had modified a congestion header. Therefore, the XCP 479 congestion header should immediately follow the IP header and precede 480 any AH. 482 3.4. NAT, middlebox issues 484 Middleboxes that attempt to perform actions invisibly on flows must 485 preserve the congestion header. Middleboxes that terminate the TCP 486 connection should terminate the XCP connection. Middleboxes that 487 insert queues into the forwarding path should participate in XCP. 489 3.5. MPLS/Tunneling Issues 491 When a flow enters an IP tunnel [RFC2003], IPsec ESP tunnel 492 [RFC2406], or MPLS [RFC3031], network ingress point, the congestion 493 header should be replicated on the "front" of the outer IP header. 494 For example, when a packet enters an IP tunnel, the following 495 transformation should occur: 497 [IP2] \_ outer header 498 __--> [XCP] / 499 [IP1]/ [IP1] \ 500 [XCP] ---> [XCP] |_ inner header 501 [TCP] [TCP] | 502 ... ... / 504 Here the XCP header appended to the front of the outer header is 505 copied from the inner header, with the appropriate change to the 506 Protocol field to indicate that the next protocol is IP. 508 When the packet exits the tunnel, the congestion header, which may 509 have been modified by routers along the tunneled path, is copied from 510 the outer header into the inner header. 512 4. XCP Functions 514 XCP is concerned with the sender and receiver end-systems and the 515 routers along the packet's path. This section describes the XCP- 516 related algorithms to be implemented in each of these entities. The 517 specific case of TCP as transport protocol is also described. The 518 emphasis in this section is on explicit and detailed definition of 519 the XCP algorithms. The theoretical derivation and analysis of the 520 algorithms can found in [Katabi03]. 522 4.1. End-System Functions 524 4.1.1. Sending Packets 526 The sender is responsible for maintaining five parameters, or their 527 equivalent: (1) a desired throughput value, (2) a current estimate of 528 the actual throughput, (3) the maximum throughput allowed by XCP, (4) 529 a current estimate of inter-packet time, denoted X, and (5) a current 530 estimate of the round-trip time 532 A sender may choose to use any reasonable value, i.e., any achievable 533 value, for desired throughput. An application may supply this value 534 via an API, or it might be the speed of the local interface. 536 When sending a packet, the sender fills in the fields of the 537 congestion header as follows: 539 o The sender sets the RTT field to a scaled smoothed round-trip time 540 estimate, or to zero if the round-trip time is not yet known. 542 o The sender sets the X field to the current inter-packet time 543 estimate, or to zero if an estimate is not yet available. Packets 544 carrying zero X field can receive negative feedback, but not 545 positive. Since XCP requires only a single round trip for a flow 546 to gain an estimate of RTT, this is expected to have negligible 547 effect. The value of X may be estimated as the smoothed round- 548 trip time estimate divided by the number of outstanding packets 549 (or congestion window size in packets, for window-based 550 protocols). Alternatively, it may be derived from the ratio of 551 the packet size to the current throughput estimate. Using 552 instantaneous RTT estimates in the calculation of X may yield 553 better results than using the smoothed RTT, especially for senders 554 with coarse-grained timers, i.e., timers with precision less than 555 the RTT. This is a subject for further study. Also,see Section 556 Section 4.1.3.3 for a discussion of RTT estimation. 558 o The sender calculates a desired change (typically, an increase) in 559 throughput. This is normally the difference between the current 560 estimated throughput and the desired throughput. However, if the 561 sender does not have sufficient data to send at the current 562 allowed throughput, the desired change in throughput SHOULD be 563 zero. 565 o The sender then divides the desired throughput change by the 566 number of packets in one round-trip time, and puts the result in 567 the Delta_Throughput field of the Congestion Header. This per- 568 packet distribution of the throughput change is necessary because 569 an XCP router does not maintain per-flow congestion state; it must 570 treat each packet independently of others in the same flow. The 571 number of packets in an RTT may be estimated by the product of the 572 current throughput and the RTT, divided by the Maximum Segment 573 Size (MSS). 575 The Delta_Throughput (in bytes/second) can be calculated as: 577 desired_throughput - Throughput 578 Delta_Throughput = ----------------------------------- 579 Throughput * ( RTT/MSS ) 581 where: 582 desired_throughput is measured in bytes/second 583 Throughput is measured in bytes/second 584 RTT is measured in seconds 585 MSS is measured in bytes 587 However, Delta_Throughput should be set to zero if, for any 588 reason, no additional capacity is needed, e.g., there is 589 insufficient data to maintain Throughput for the next RTT, as 590 discussed above. 592 o An issue for future consideration is how to treat the case when 593 Delta_Throughput is calculated to be < 1 Bps. Since an integer 594 representation is passed in the Congestion Header, the result will 595 appear as zero. It would be possible to send a fraction of the 596 packets in a round trip time with non-zero Delta_Throughput 597 values. 599 o For TCP, the throughput estimate can be obtained by dividing the 600 congestion window cwnd (in bytes) by RTT (in seconds), for 601 example. Alternatively, it could be measured. The ratio cwnd/RTT 602 differs from true throughput in two respects. First, cwnd doesn't 603 account for header size. This may become significant should XCP 604 be applied to real-time flows that send large numbers of small 605 packets, but it is probably not much worry for TCP flows that tend 606 to use the largest possible packet size. Second, cwnd represents 607 permission for the sender to transmit data. If the application 608 doesn't use all of the available cwnd, the advertised throughput 609 will be larger than the true throughput. This may result in an 610 XCP router creating an unfair allocation of negative feedback to a 611 flow. 613 4.1.2. Processing Feedback at the Receiver 615 An XCP receiver is responsible for copying the Delta_Throughput data 616 it sees on arriving packets into the Reverse_Feedback field of 617 outgoing packets. In TCP, outgoing packets would normally be ACK- 618 only segments. 620 In some cases returning packets are sent less frequently than 621 arriving packets, e.g., with delayed acknowledgments [RFC1122]. The 622 receiver is responsible for calculating the sum of the arriving 623 Delta_Throughput fields for placement in outgoing Reverse_Feedback 624 fields. 626 4.1.2.1. Feedback from the Receiver 628 The receiver end-system returns XCP congestion feedback from the 629 network to the sender, by copying the Delta_Throughput information 630 from arriving packets into the Reverse_Feedback field of Congestion 631 Headers in outgoing packets (possibly aggregating data as described 632 in Section 4.1.2). 634 It is possible that even empty ACK packets may create or encounter 635 congestion in the reverse direction. Although TCP implementations 636 generally do not perform congestion-based pacing of empty ACK 637 segments, some transport protocols (e.g., DCCP) may be. Such a 638 transport protocol may choose to use XCP congestion control on the 639 returning ACKs as well as on the data. 641 In the normal case of a unidirectional data flow with XCP applied 642 only to that data flow, the feedback can be sent in a Minimal format 643 Congestion Header, in which the RTT, X, and Delta_Throughput fields 644 are set to zero. 646 4.1.3. Processing Feedback at the Sender 648 When packets arrive back to the sender carrying reverse feedback, the 649 sender must adjust its sending rate accordingly. 651 As noted earlier, this throughput adjustment may be made in TCP by 652 updating the sender's congestion window, cwnd. This should use the 653 formula: 655 cwnd = max(cwnd + feedback * RTT, MSS) 657 where: 658 cwnd = current congestion window (bytes) 659 feedback = Reverse_Feedback field from received packet, 660 (bytes/sec, may be +/-) 661 RTT = Sender's current round-trip time estimate 662 (seconds) 663 MSS = maximum segment size (bytes) 665 The value of cwnd has a minimum of MSS to avoid the "Silly Window 666 Syndrome" [RFC0813]. 668 4.1.3.1. Aging the Allowed Throughput 670 When a sending application does not send data fast enough to fully 671 utilize the allowed throughput, XCP should reduce the allowed 672 throughput as time passes, to avoid sudden bursts of data into the 673 network if the application starts to send data later. 675 We present a slight modification of the algorithm for aging the 676 allowed throughput below. It is based on Section 4.5 of [Katabi03]. 677 Each RTT in which the sender sends with actual throughput which is 678 less than the allowed throughput, the allowed throughput MUST be 679 reduced by the following exponential averaging formula: 681 Allowed_Throughput = Allowed_Throughput*(1-p) + 682 Actual_Throughput * p 684 where: p is a parameter controlling the speed of aging, 685 ranged between 0 and 1. 687 Using p = 0.5 is suggested. Consideration of values of p or other 688 algorithms is a research topic. 690 4.1.3.2. Response to Packet Loss 692 When the transport protocol is TCP, a packet drop or detection of an 693 ECN notification [RFC3168] should trigger a transition to standard 694 TCP congestion control behavior[RFC2581]. In other words, cwnd 695 should be halved and Jacobson's fast retransmission/fast recovery, 696 slow start, and congestion avoidance algorithms should be applied for 697 the remainder of the connection or until the congestion event is 698 known to have passed (see Section 5.1 for discussion of alternative 699 approaches to this issue. The assumption is that the packet drop 700 reveals the presence of a congested non-XCP router in the path. 701 Transitioning to standard TCP behavior is a conservative response. 703 Note also the following: 705 o The change in congestion control algorithm should be delayed until 706 the three DUPACKs have arrived, according to the Fast 707 Retransmission/Fast Recovery algorithm[RFC2581]. 709 o Once the change to standard TCP congestion control has occurred, 710 cwnd should be managed using the RFC2581 algorithm. 712 o The X field in outgoing packets should continue to reflect the 713 current inter-packet time. This allows the XCP processes in the 714 routers along the path to continue to monitor the flow's 715 utilization. 717 o Further study is needed to determine whether it will be possible 718 to return a connection to XCP congestion control, once it has 719 transitioned to Van Jacobson mode. 721 o For transport protocols other than TCP, the response to a packet 722 loss or ECN notification is a subject for further study. 724 4.1.3.3. RTT Estimates 726 Having a good estimate of the round trip time is more important in 727 XCP than in Van Jacobson congestion control. There is evidence that 728 small errors in the RTT estimate can result in larger errors in the 729 throughput and X estimates. The current cwnd divided by SRTT is only 730 an approximation of the actual throughput. Likewise, SRTT divided by 731 cwnd in packets is only an approximation of the highly variable 732 inter-packet time, X. The RTT used in the ns-2 code in [KHR02] used a 733 smoothed floating-point RTT estimator, rather than instantaneous 734 measurements. Additional research is needed to develop 735 recommendations for RTT estimation. 737 4.2. Router functions 739 The router calculations for XCP are divided into those that occur 740 upon packet arrival, those that occur upon control interval timeout, 741 those that occur upon packet departure, and the assessment of the 742 persistent queue, which uses a separate timer. The calculations are 743 presented in the following sections as annotated pseudo-code. 745 4.2.1. Calculations Upon Packet Arrival 747 When a packet arrives at a router, several parameters used by XCP 748 need to be updated. The steps are described in the following pseudo- 749 code. 751 ======================================================== 752 On packet arrival do: 754 1. input_traffic += Pkt_size 756 2. sum_x += X 758 3. if (Rtt < MAX_INTERVAL) then 760 4. sum_xrtt += X * Rtt 762 5. else 764 6. sum_xrtt += X * MAX_INTERVAL 765 ======================================================== 767 Line 1: The variable input_traffic accumulates the volume of data 768 that have arrived during a control interval. When a packet 769 arrives, the packet size is taken from the IP header and is added 770 to the ongoing count. 772 Line 2: The variable sum_x is used in the control interval 773 calculation (see equation 4.2 of [Katabi03]) and in capacity 774 allocation. For each packet, values of X from the XCP header is 775 accumulated. It is recommended that sum_x is stored in a 64-bit 776 unsigned integer variable. 778 Lines 3 and 5: A test is performed to check whether the round trip 779 time of the flow exceeds the maximum allowable control interval. 780 If so, MAX_INTERVAL, the maximum allowable control interval, is 781 used in the subsequent calculations. Too large a control interval 782 will delay new flows from acquiring their fair allocation of 783 capacity. See Section 4.2.4 for a discussion of the recommended 784 value for MAX_INTERVAL. 786 Lines 4 and 6: As in Line 2, the variable sum_xrtt is used in the 787 control interval calculation. It is recommended that it is stored 788 in a 96-bit unsigned variable. 790 4.2.2. Calculations Upon Control Interval Timeout 792 When the control timer expires, several variables need to be updated 793 as shown below. 795 Note that several calculations show divisions. These divisions 796 should either be accomplished using floating-point arithmetic or 797 integer arithmetic and appropriate scaling to avoid over- or under- 798 flow. 800 ======================================================== On 801 estimation-control timeout do: 803 7. avg_rtt = sum_xrtt / sum_x 805 8. input_bw = input_traffic / ctl_interval 807 9. F = a * (capacity - input_bw) - b * queue / avg_rtt 809 10. shuffled_traffic = shuffle_function(...) 811 11. residue_pos_fbk = shuffled_traffic + max(F,0) 813 12. residue_neg_fbk = shuffled_traffic + max(-F,0) 815 13. Cp = residue_pos_fbk / sum_x 817 14. Cn = residue_neg_fbk / input_traffic 819 15. input_traffic = 0 821 16. sum_x = 0 823 17. sum_xrtt = 0 825 18. ctl_interval = max(avg_rtt, MIN_INTERVAL) 827 19. timer.reschedule(ctl_interval) 828 ======================================================== 829 Line 7: Update avg_rtt by taking the ratio of the two sums 830 accumulated in the previous section. This value is used to 831 determine the control interval (line 17). 833 Line 8: The average bandwidth of arriving traffic is calculated by 834 dividing the bytes received in the previous control interval by 835 the duration of the previous control interval. 837 Line 9: The aggregate feedback, F, is calculated. The variable 838 'capacity' is the ability of the outbound link to carry IP 839 packets, in bytes/second. The variable 'avg_rtt' was calculated 840 in line 7. The variable 'queue' is the persistent queue and is 841 defined in section Section 4.2.5. The values a and b are constant 842 parameters. According to [Katabi03], the constant a may be any 843 positive number such that a < (pi/4*sqrt(2)). A nominal value of 844 0.4 is recommended. The constant b is defined to be b = 845 a^2*sqrt(2). (If the nominal value of a is used, the value for b 846 would be 0.226.) Note that F may be positive or negative. 848 Line 10: This line establishes the amount of capacity that will be 849 shuffled in the next control interval through the use of the 850 shuffle_function. Shuffling takes a small amount of the available 851 capacity and redistributes it by adding it to both the positive 852 and negative feedback pools. This allows new flows to acquire 853 capacity in a full loaded system. 855 The recommended shuffle_function is as follows: 857 shuffled_traffic = max(0, 0.1 * input_bw - |F|) 859 The variable 'input_bw' is defined above in Line 8. Implementers 860 may choose other functions. It is important to consider that more 861 shuffled traffic decreases the time for new flows to acquire 862 capacity and converge to fairness. However, too much shuffling 863 may impede flows from acquiring their fair share of available 864 capacity. (For example, consider a setup of N flows bottlenecked 865 downstream from the given router and another flow, not limited as 866 those, trying to acquire its fair share. In this case shuffling 867 leads to under-utilization of the available bandwidth and impedes 868 the unlimited flow.) Shuffled_traffic is always a positive value. 870 The objective of the feedback calculations is to obtain a per-packet 871 feedback allocation from the router. Lines 13 and 14 obtain factors 872 in this calculation that, unfortunately, have no physical meaning. 873 One might view them as per-flow capacity allocations that have some 874 additional processing to prepare them for per-packet allocation. 875 Note that, with the use of shuffled_traffic, a non-idle router will 876 always start a control interval with non-zero values for both Cn and 877 Cp. 879 Line 11: The variable 'residue_pos_fbk' keeps track of the pool of 880 available positive capacity a router has to allocate. It is 881 initialized to the positive aggregate feedback. 883 Line 12: The variable 'residue_neg_fbk' keeps track of the pool of 884 available negative capacity a router has to allocate. It is 885 initialized to the negative aggregate feedback. This variable is 886 always positive. 888 Line 13: This line calculates the positive feedback scale factor, Cp. 889 The variables residue_pos_fbk, and sum_x are defined above. 891 Line 14: This line calculates the negative feedback scale factor, Cn. 892 This is a positive value. The definitions for residue_neg_fbk, 893 and input_traffic are given above. 895 Line 15-17: Reset various counters for the next control interval. 897 Line 18: Set the next control interval. The use of MIN_INTERVAL is 898 important to establish a reasonable control interval when the 899 router is idle. 901 Line 19: Set timer. 903 4.2.3. Calculations Upon Packet Departure 905 An XCP router processes each packet using the feedback parameters 906 calculated above. As stated earlier, each packet indicates the 907 current inter-packet time (X) and a throughput adjustment, 908 Delta_Throughput. The router calculates a per-packet capacity change 909 which will be compared to the Delta_Throughput field in the packet 910 header. Using the AIMD rule, positive feedback is applied equally 911 per-flow, while negative feedback is made proportional to each flow's 912 capacity. 914 To accommodate high-speed routers, XCP uses a fixed-point numeric 915 representation for the Congestion Header fields. This means that the 916 per-packet calculations defined below result in residual error that 917 is less than 1 Bps per packet. These errors accumulate across all 918 the packets in a control interval, resulting in an inaccuracy in 919 XCP's allocation of available bandwidth to flows. Further work is 920 needed to understand whether this will be a significant problem and, 921 if so, whether there is any solution short of using 64 bit precision 922 or floating point. 924 Processing should be done according to the pseudo-code below. 926 ======================================================== 927 On packet departure: 929 20. pos_fbk = Cp * X 931 21. neg_fbk = Cn * Pkt_size 933 22. feedback = pos_fbk - neg_fbk 935 23. if(Delta_Throughput > feedback) then 937 24. Delta_Throughput = feedback 939 25. else 941 26. neg_fbk = min(residue_neg_fbk, neg_fbk + 942 (feedback - Delta_Throughput)) 944 27. pos_fbk = Delta_Throughput + neg_fbk 946 28. residue_pos_fbk = max(0, residue_pos_fbk - pos_fbk) 948 29. residue_neg_fbk = max(0, residue_neg_fbk - neg_fbk) 950 30. if (residue_pos_fbk <= 0) then Cp = 0 952 31. if (residue_neg_fbk <= 0) then Cn = 0 954 ======================================================== 956 Line 20: The contribution of positive feedback for the current packet 957 is calculated using Cp, defined in line 13, and X (the flow's 958 advertised inter-packet time) from the Congestion Header. Note 959 that if Cp (and Cn in Line 21) is implemented as a floating point 960 number, this calculation would be implemented by multiplying the 961 Cp-mantissa by the value of X, then shifting the result by the 962 amount of the Cp-exponent. 964 Line 21: The contribution of negative feedback for the current packet 965 is calculated using Cn, defined in line 14, and Pkt_size from the 966 IP header. This value of neg_fbk is positive. 968 Line 22: The router's allocated feedback for the packet is the 969 positive per-packet feedback minus the negative per-packet 970 feedback. This value may be positive or negative. 972 Line 23-24: Line 23 tests whether the packet is requesting greater 973 capacity increase (via the packet's Delta_Throughput field) than 974 the router has allocated. If so, this means the the sender's 975 desired throughput needs to be reduced to be the router's 976 allocation. In line 24 the Delta_Throughput field in the packet 977 header updated with the router feedback allocation. 979 Line 25: This branch is executed when the packet is requesting a 980 smaller throughput increase than the router's allocation. In this 981 branch, and the rest of this pseudo-code, the packet header is not 982 updated and the remaining code is to correctly update the feedback 983 pool variables. 985 Line 26: In this line, the packet's negative feedback contribution, 986 neg_fbk, is set to be the smaller of two terms. The first term, 987 residue_neg_fbk, is the pool of negative feedback, i.e., this 988 drains the remaining negative feedback in the pool. The second 989 term increases the nominal negative feedback from the router by 990 the amount which the Delta_Throughput is less than net router 991 allocation. This allows the router to capture feedback which is 992 allocated by an upstream bottleneck. 994 Line 27: The positive allocation, pos_fbk, is adjusted to be the sum 995 of Delta_Throughput and neg_fbk, from Line 26. This is required 996 for the sum of pos_fbk and neg_fbk to equal Delta_Throughput. 998 Line 28-29: In these two lines, the feedback pools, residue_pos_fbk 999 and residue_neg_fbk, are reduced by the values of pos_fbk and 1000 neg_fbk accordingly, but prevented from going negative. 1002 Line 30-31: When a feedback pool becomes empty, set the scale factor 1003 to zero, i.e., stop handing out associated feedback. 1005 4.2.4. The Control Interval 1007 The capacity allocation algorithm in XCP router updates several 1008 parameters every Control Interval. The Control Interval is currently 1009 defined to be the average RTT of the flows passing through the 1010 router, i.e., avg_rtt calculated in Line 7 above. Other possible 1011 choices for the control interval are under study. 1013 Notes on avg_rtt: 1015 o In this document, the quantity 'avg_rtt' refers to the last 1016 calculated value. In other words, the avg_rtt calculated based on 1017 packets arriving in the previous control interval. 1019 o The avg_rtt calculation should ignore packets with an RTT of zero 1020 in the header. 1022 o avg_rtt MUST have a minimum value. This is to allow flows to 1023 acquire bandwidth from a previously idle router. The default 1024 minimum value, MIN_INTERVAL, should be max(5-10ms, propagation 1025 delay on attached link). 1027 o avg_rtt MUST have a maximum value. The default maximum value, 1028 MAX_INTERVAL, should be max(0.5-1 sec, propagation delay on 1029 attached link). 1031 4.2.5. Obtaining the Persistent Queue 1033 In Section 4.2.2 the variable 'queue' contains the persistent queue 1034 over the control interval. This is intended to be the minimum 1035 standing queue over the queue estimation interval. 1037 The following pseudo-code describes how to obtain the minimum 1038 persistent queue: 1040 ======================================================== 1041 On packet departure do: 1043 32. min_queue = min(min_queue, inst_queue) 1045 ======================================================== 1047 When the queue-computation timer expires do: 1049 33. queue = min_queue 1051 34. min_queue = inst_queue 1053 35. Tq = max(ALLOWED_QUEUE, (avg_rtt - inst_queue/capacity)/2) 1055 36. queue_timer.reschedule(Tq) 1057 ======================================================== 1059 Line 32: The current instantaneous queue length is checked each time 1060 a packet departs compute the minimum queue size. 1062 If avg_rtt is being used as the Control Interval, it MUST NOT be used 1063 as the interval for measuring the minimum persistent queue. Doing so 1064 can result in a feed-forward loop. For example, if a queue develops 1065 the average RTT will increase. If the avg_rtt increases, it takes 1066 longer to react to the growing queue and the queue gets larger, 1067 leading to instability. 1069 Line 33: Upon expiration of the queue estimation timer, Tq, the 1070 variable queue, the persistent queue, is set to be the minimum 1071 queue occupancy over the last Tq. 1073 Line 34: Upon expiration of the queue estimation timer, reset the 1074 running estimate of the minimum queue to be the current queue 1075 occupancy. 1077 Line 35: The first term in the max function, ALLOWED_QUEUE, is the 1078 time to drain a standing queue that you are willing to tolerate. 1079 (A nominal value of 2ms worth of queuing is recommended but this 1080 may be tuned by implementers.) The second term is an estimate of 1081 the propagation delay. In other words the persistent queue is a 1082 queue that does not drain in a propagation delay. the division by 1083 2 is a conservative factor to avoid overestimating the propagation 1084 delay. 1086 Line 36: The queue computation timer is set. 1088 5. Unresolved Issues 1090 XCP is a work-in-progress. This section describes some known issues 1091 that need to be resolved. 1093 5.1. XCP With Non-XCP Routers 1095 Obviously, non-XCP routers will exist in networks before XCP becomes 1096 ubiquitously deployed and we expect other non-XCP systems to continue 1097 to be in the network indefinitely. Long term non-XCP network 1098 elements include any sort of link-level switches with queuing, e.g. 1099 ATM switches and sophisticated Ethernet switches. Even simple 1100 multiplexers are non-XCP queues with very little buffering. 1102 Sources and the network care about these non-XCP elements because any 1103 one of them can be a site of network congestion, and if an XCP 1104 endpoint is bottlenecked at one of these non-XCP elements, no router 1105 feedback will inform the endpoint to slow down. If nothing is done, 1106 such an element will probably collapse under congestion. 1108 Although exactly how XCP sources will operate in this environment is 1109 an open issue, a current promising direction is for endpoints to run 1110 a traditional end-to-end congestion detection algorithm in parallel 1111 with the XCP algorithm and switch over to using that algorithm for 1112 control when congestion is detected that XCP is not controlling. For 1113 example, an XCP source that detects 3 duplicate acknowledgments would 1114 fall back to TCP Reno behavior. 1116 An endpoint that is limited by its end-to-end congestion algorithm 1117 would indicate so to XCP routers by setting a bit in the packet 1118 header. A router may process such packets differently than packets 1119 from endpoints that are being controlled by XCP. For example, the 1120 router might allocate end-to-end controlled packets less feedback or 1121 not reduce its feedback pools by the full amount when assigning 1122 feedback to those packets. 1124 Though using its end-to-end algorithm to control its sending rate, an 1125 endpoint will also monitor the XCP feedback and if the source 1126 discovers that the XCP feedback would be more restrictive than the 1127 end-to-end control over a round trip time, the endpoint will revert 1128 to following XCP feedback. XCP feedback that is more restrictive 1129 over a round trip time is an indication that the endpoint's 1130 bottleneck is once again at an XCP router and the endpoint should 1131 take advantage of the more precise XCP information. 1133 Evaluation of these algorithms is ongoing work 1135 5.2. Variable Rate Links 1137 As discussed in [Zhang05], XCP may perform poorly over shared links. 1138 When a link is shared, such as in CSMA ethernet or 802.11 wireless 1139 networks, a single queue's drain rate is often a function of the load 1140 in the shared medium. So, using a constant value for the variable 1141 'capacity' in the routing control algorithm may not work well. For 1142 correct operation, the XCP router's notion of capacity needs to 1143 reflect how the link capacity is shared. 1145 5.3. XCP as a TCP PEP 1147 In the Internet today TCP performance-enhancing proxies (PEPs) are 1148 sometimes used to improve application performance over certain 1149 networks. TCP PEPs, and the issues surrounding their use are 1150 described in [RFC3135]. A common mechanism used in TCP PEPs is to 1151 split a TCP connection into three parts where the first and last run 1152 TCP and a more aggressive transport protocol is run in the middle 1153 (across the path which generates poor TCP performance). This 1154 improves performance over the "problematic" portion of the path and 1155 does not require changing the protocol stacks on the end systems. 1156 For example, if a high-speed satellite link was used to connect a LAN 1157 to the Internet, a TCP PEP may be placed on either side of the 1158 satellite link. 1160 It is not unusual today to find TCP PEPs which, to get high data 1161 rates, do not use congestion control at all. Of course, this limits 1162 the environments in which they can be used. However, XCP may be used 1163 in between two TCP PEPs to get high transfer rates and still respond 1164 to congestion in a correct and scalable way. 1166 Work on using XCP as a TCP PEP is just beginning [Kapoor05]. 1167 Objectives for such a mechanism would be: 1169 o preserve end-to-end TCP semantics as much as possible 1171 o enable some form of discovery for one PEP to determine that 1172 another PEP was in the path 1174 o allow for recovery should one half of a PEP pair fail or the route 1175 change so that one or both PEPs are not on the path 1177 o enable aggregation of multiple flows between two PEPs. 1179 A system which met the above objectives could also be used for 1180 incremental deployment of XCP. A network operator could deploy XCP- 1181 capable routers and use PEPs at the periphery of the network to 1182 convert from traditional TCP to XCP congestion control. This may 1183 result in smaller queues and improved link utilization within the XCP 1184 network. (This may be of more interest to wireless network providers 1185 than to over-provisioned fiber backbones.) 1187 5.4. Sharing resources between XCP and TCP 1189 Katabi describes a system for sharing router bandwidth between XCP 1190 and TCP flows that is based on sharing the output capacity based on 1191 the average throughputs of XCP and TCP sources using the 1192 router.[Katabi03] Two queues are installed at the router and they are 1193 served with different weights by the forwarding engine. The 1194 algorithm is work-conserving; the forwarding engine is never idle 1195 when either queue has packets in it. 1197 A TFRC-like algorithm estimates the average congestion window of TCP 1198 sources and the XCP algorithm estimates the average throughput of XCP 1199 sources. These averages are used to dynamically weight the time the 1200 processor spends on each queue. Initial experiments indicate that 1201 this system can provide fairness between the two flow classes (XCP/ 1202 TCP).[Katabi03] 1204 Open issues remain, however; for example there are questions about 1205 how well the TFRC-like algorithm can estimate TCP throughput with 1206 only access to local drop rates, convergence time of the weighting 1207 algorithm has never been explored, and no system for buffer 1208 allocation to complement the link capacity allocation has been put 1209 forward. These open issues are under study. 1211 5.5. A Generalized Router Model 1213 The XCP algorithm described here and in [Katabi03] manages congestion 1214 at a single point in a router, most likely an output queue. However, 1215 resource contention can occur at many points in a router. Input 1216 queues, backplanes, computational resources can 'congest' in addition 1217 to output buffers. There is a need to develop a general model and a 1218 variety of mechanisms to identify and manage resource contention 1219 throughout the router. 1221 5.6. Host back-to-back operation 1223 XCP hosts should be capable of back-to-back operation, i.e., with no 1224 router in the path. Nominally, this should not be a problem. A 1225 sender initializes delta_throughput to the desired value, no router 1226 modifies it and, thus, it is automatically granted. However, it has 1227 not yet been decided whether an XCP receiver should be capable of (or 1228 require) adjusting the delta_throughput to request flow control from 1229 the receiver to the sender. 1231 At this point, XCP offers no mechanism for flow control. (Open 1232 question: Should it?) It is believed that running XCP on the output 1233 queue of a host would solve this problem. However, it isn't clear 1234 that the complexity is justified by the need to solve this situation. 1236 6. Security Considerations 1238 The presence of a header which may be read and written by entities 1239 not participating in the end-to-end communication opens some 1240 potential security vulnerabilities. This section describes them and 1241 tries to give enough context so that users can understand the risks. 1243 Man-in-the-Middle Attacks 1245 There is a man-in-the-middle attack where a malicious user can 1246 force a sender to stop sending by inserting negative feedback into 1247 flow. This is little different from a malicious user discarding 1248 packets belonging to a flow using VJ congestion control or setting 1249 ECN bits. One question worth investigating further is whether the 1250 XCP attack is harder to diagnose. 1252 Covert Data Channels 1254 IPsec needs to be modified, as discussed in Section 3.3, to allow 1255 routers to read the entire congestion header and write the 1256 delta_feedback field. This could become a covert data-channel, 1257 i.e., a way in which an end-system can make data viewable to 1258 observers in the network, on a compromised end-system. 1260 Malicious Sources 1262 The XCP algorithms rely on senders to advertise information about 1263 their current RTT and X and correctly respond to feedback 1264 delivered from the network. Naturally, the possibility occurs 1265 that a sender won't perform these functions correctly. Chapter 7 1266 of [Katabi03] and [Katabi04] examine these issues. 1268 A source which lies about its values of X and hence throughput 1269 cannot affect the link utilization and, in the worst case, can 1270 unfairly acquire capacity. However, this is equivalent to a 1271 sender opening up multiple TCP flows. So, there is an incentive 1272 to lie about X. However, because X is explicitly stated in each 1273 packet header, it is a simpler matter to police it at the edge of 1274 the network than, say, for TCP. 1276 A source which lies about its RTT can disrupt the router control 1277 algorithm, particularly when a large number of sources lie about 1278 their RTT and the router control interval is adaptive and uses the 1279 average RTT. However, there is little incentive to lie as it will 1280 not affect the fair allocation of capacity and the liar will 1281 experience the same degradation as the non-lying flows. Lying 1282 about RTT should be considered a weak denial-of-service attack. 1284 A flow may also ignore negative feedback from the router. Such a 1285 flow can obtain unfair throughput in a congested router. However, 1286 as with lying, the explicit nature of XCP makes it possible to 1287 verify that flows are responding to feedback. For example, a 1288 policing function in the path (presumeably near the edge so that 1289 the load is manageable and it can be expected to see packet flow 1290 in both directions) may inspect congestion headers for a flow in 1291 both directions. If the policer sees negative feedback heading 1292 towards a source and no reduction in throughput it may, e.g., 1293 punish the flow by severely restricting the throughput. Note that 1294 this can be applied on a probabilistic basis, sampling flows only 1295 occasionally. 1297 7. IANA Considerations 1299 XCP requires the assignment of an IP protocol number. Once this 1300 value has been assigned, the number may be inserted (by the RFC 1301 Editor) into Section 3.1 and this paragraph may be removed prior to 1302 publication. 1304 8. Acknowledgements 1306 The authors would like to acknowledge the many contributors who have 1307 assisted in this work. Bob Braden applied his usual sage guidance to 1308 the project and to the spec, in particular. Ted Faber wrote the 1309 initial implementation framework and provided much wisdom on kernel 1310 development and congestion control. John Wroclawski advised on 1311 project priorities and strategy. Eric Coe developed the initial 1312 implementation and testbed. Aman Kapoor performed supporting 1313 simulations and debugged kernel code. Padma Haldar ported ns-2 1314 simulation code to ns-2 distribution. Jasmeet Bagga and Anuraag 1315 Mittal conducted simulations on various aspects of XCP performance. 1316 On the XCP mailing list, Tim Shepherd, Tom Henderson, and Matt Mathis 1317 made valuable contributions to the effort. To all the above go our 1318 sincere thanks. 1320 9. References 1322 [I-D.ietf-dccp-spec] 1323 Kohler, E., "Datagram Congestion Control Protocol (DCCP)", 1324 draft-ietf-dccp-spec-04 (work in progress), July 2003. 1326 [Jacobson88] 1327 Jacobson, V., "Congestion Avoidance and Control", ACM 1328 Computer Communication Review Proceedings of the Sigcomm 1329 '88 Symposium, August 1988. 1331 [KHR02] Katabi, D., Handley, M., and C. Rohr, "Internet Congestion 1332 Control for Future High Bandwidth-Delay Product 1333 Environments", ACM Computer Communication 1334 Review Proceedings of the Sigcomm '02 Symposium, 1335 August 2002. 1337 [Kapoor05] 1338 Kapoor, A., Falk, A., Faber, T., and Y. Pryadkin, 1339 "Achieving Faster Access to Satellite Link Bandwidth", 1340 IEEE 8th IEEE Global Internet Symposium, Miami, FL, March 1341 2005, 2005. 1343 [Katabi03] 1344 Katabi, D., "Decoupling Congestion Control and Bandwidth 1345 Allocation Policy With Application to High Bandwidth-Delay 1346 Product Networks", MIT PhD. Thesis, March 2003. 1348 [Katabi04] 1349 Katabi, D., "XCP's Performance in the Presence of 1350 Malicious Flows", Second International Workshop on 1351 Protocols for Fast Long-Distance Networks, Presentation, 1352 February 2004. 1354 [Padhye98] 1355 Padhye, J., Firoiu, V., Towsley, D., and J. Krusoe, 1356 "Modeling TCP throughput: A simple model and its empirical 1357 validation", ACM SIGCOMM '98 conference on 1358 Applications,technologies, architectures, and protocols 1359 for computer communication, 1998. 1361 [RFC0791] Postel, J., "Internet Protocol", STD 5, RFC 791, 1362 September 1981. 1364 [RFC0793] Postel, J., "Transmission Control Protocol", STD 7, 1365 RFC 793, September 1981. 1367 [RFC0813] Clark, D., "Window and Acknowledgement Strategy in TCP", 1368 RFC 813, July 1982. 1370 [RFC1122] Braden, R., "Requirements for Internet Hosts - 1371 Communication Layers", STD 3, RFC 1122, October 1989. 1373 [RFC2003] Perkins, C., "IP Encapsulation within IP", RFC 2003, 1374 October 1996. 1376 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 1377 Requirement Levels", BCP 14, RFC 2119, March 1997. 1379 [RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, 1380 S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., 1381 Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, 1382 S., Wroclawski, J., and L. Zhang, "Recommendations on 1383 Queue Management and Congestion Avoidance in the 1384 Internet", RFC 2309, April 1998. 1386 [RFC2401] Kent, S. and R. Atkinson, "Security Architecture for the 1387 Internet Protocol", RFC 2401, November 1998. 1389 [RFC2402] Kent, S. and R. Atkinson, "IP Authentication Header", 1390 RFC 2402, November 1998. 1392 [RFC2406] Kent, S. and R. Atkinson, "IP Encapsulating Security 1393 Payload (ESP)", RFC 2406, November 1998. 1395 [RFC2581] Allman, M., Paxson, V., and W. Stevens, "TCP Congestion 1396 Control", RFC 2581, April 1999. 1398 [RFC2960] Stewart, R., Xie, Q., Morneault, K., Sharp, C., 1399 Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M., 1400 Zhang, L., and V. Paxson, "Stream Control Transmission 1401 Protocol", RFC 2960, October 2000. 1403 [RFC3031] Rosen, E., Viswanathan, A., and R. Callon, "Multiprotocol 1404 Label Switching Architecture", RFC 3031, January 2001. 1406 [RFC3135] Border, J., Kojo, M., Griner, J., Montenegro, G., and Z. 1407 Shelby, "Performance Enhancing Proxies Intended to 1408 Mitigate Link-Related Degradations", RFC 3135, June 2001. 1410 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 1411 of Explicit Congestion Notification (ECN) to IP", 1412 RFC 3168, September 2001. 1414 [Zhang05] Zhang, Y. and T. Henderson, "An Implementation and 1415 Experimental Study of the eXplicit Control Protocol 1416 (XCP)", IEEE Proceedings of the 24th IEEE International 1417 Conference on Computer Communications (INFOCOM 2005), pp. 1418 1037-1048, Miami, Florida, Mar 2005., 2005. 1420 Authors' Addresses 1422 Aaron Falk 1423 USC Information Sciences Institute 1424 4676 Admiralty Way 1425 Suite 1001 1426 Marina Del Rey, CA 90292 1428 Phone: 310-448-9327 1429 Email: falk@isi.edu 1430 URI: http://www.isi.edu/~falk 1432 Yuri Pryadkin 1433 USC Information Sciences Institute 1434 4676 Admiralty Way 1435 Suite 1001 1436 Marina Del Rey, CA 90292 1438 Phone: 310-448-8417 1439 Email: yuri@isi.edu 1441 Dina Katabi 1442 Massachusetts Institute of Technology 1443 200 Technology Square 1444 Cambridge, MA 02139 1446 Phone: 617-324-6027 1447 Email: dk@mit.edu 1448 URI: http://www.ana.lcs.mit.edu/dina/ 1450 Intellectual Property Statement 1452 The IETF takes no position regarding the validity or scope of any 1453 Intellectual Property Rights or other rights that might be claimed to 1454 pertain to the implementation or use of the technology described in 1455 this document or the extent to which any license under such rights 1456 might or might not be available; nor does it represent that it has 1457 made any independent effort to identify any such rights. Information 1458 on the procedures with respect to rights in RFC documents can be 1459 found in BCP 78 and BCP 79. 1461 Copies of IPR disclosures made to the IETF Secretariat and any 1462 assurances of licenses to be made available, or the result of an 1463 attempt made to obtain a general license or permission for the use of 1464 such proprietary rights by implementers or users of this 1465 specification can be obtained from the IETF on-line IPR repository at 1466 http://www.ietf.org/ipr. 1468 The IETF invites any interested party to bring to its attention any 1469 copyrights, patents or patent applications, or other proprietary 1470 rights that may cover technology that may be required to implement 1471 this standard. Please address the information to the IETF at 1472 ietf-ipr@ietf.org. 1474 Disclaimer of Validity 1476 This document and the information contained herein are provided on an 1477 "AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS 1478 OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET 1479 ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED, 1480 INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE 1481 INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED 1482 WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. 1484 Copyright Statement 1486 Copyright (C) The Internet Society (2006). This document is subject 1487 to the rights, licenses and restrictions contained in BCP 78, and 1488 except as set forth therein, the authors retain all their rights. 1490 Acknowledgment 1492 Funding for the RFC Editor function is currently provided by the 1493 Internet Society.